Fix for [Bug]: KeyError in DoubleMLPLIV.fit() with multiple instruments and store_predictions=True

[vnastl]

Description

In the case of multiple instruments, the function DoubleMLPLIV.fit() throws an error when executed with the parameter 'store_predictions=True', as reported in [Bug]: KeyError in DoubleMLPLIV.fit() with multiple instruments and store_predictions=True.

This occurs because the list of learner names is defined as follows in the case of multiple instruments:

line 269 in double_ml_pliv.py: param_names = ['ml_l', 'ml_r'] + ['ml_m_' + z_col for z_col in self._dml_data.z_cols]

The predictions for the nuisance functions are however saved using the keys 'ml_l', 'ml_m', 'ml_r', 'ml_g'.

line 397 - 405 in double_ml_pliv.py: 
        preds = {'predictions': {'ml_l': l_hat['preds'],
                                 'ml_m': m_hat['preds'],
                                 'ml_r': r_hat['preds'],
                                 'ml_g': g_hat['preds']},
                 'models': {'ml_l': l_hat['models'],
                            'ml_m': m_hat['models'],
                            'ml_r': r_hat['models'],
                            'ml_g': g_hat['models']}
                 }

To store the predictions if store_predictions=True in DoubleMLPLIV.fit(), we use preds['predictions'] but iterate over param_names. Hence, the reported KeyError occurred.

To solve this error, I propose to add the predictions and models made for each instrument in the case of multiple instruments, i.e. for the list of learners

['ml_m_' + z_col for z_col in self._dml_data.z_cols],

to the dictionaries preds['predictions'] and preds['models'], respectively.
The proposed code passes all (unit) tests.

Reference to Issues or PRs

[Bug]: KeyError in DoubleMLPLIV.fit() with multiple instruments and store_predictions=True #184

[SvenKlaassen]

Thanks again, @vnastl. This problem should be solved with #182. Therefore, I will close this pull request.